NVIDIA Blackwell Ultra B300 SXM

Description

NVIDIA Blackwell Ultra B300 is the SXM GPU generation used on official NVIDIA HGX B300 baseboards and in the NVIDIA DGX B300 system. Public documentation describes an eight-GPU HGX board, not a standalone add-in card. The figures below follow NVIDIA’s published HGX B300 table rather than unofficial per-GPU estimates.

NVIDIA lists the HGX B300 configuration as eight NVIDIA Blackwell Ultra SXM GPUs with 2.1 TB of total GPU memory, fifth-generation NVLink and NVLink Switch, 1.8 TB/s of GPU-to-GPU bandwidth, and 14.4 TB/s of aggregate NVLink bandwidth. At the eight-GPU board level, NVIDIA publishes FP4 Tensor performance of 144 PFLOPS with sparsity and 108 PFLOPS dense, 72 PFLOPS of sparse FP8/FP6 Tensor performance (dense is half of the sparse figure), and 600 TFLOPS of FP32. Networking bandwidth is listed at 1.6 TB/s, and attention performance is described as 2x versus NVIDIA Blackwell. The official HGX page does not publish a per-GPU memory capacity in gigabytes or a per-GPU TDP for B300; those values are therefore omitted here.

DGX B300 is documented as an air-cooled, MGX-compliant eight-GPU system using Blackwell Ultra SXM modules. OEM hosts such as the GIGABYTE G894-SD3 implement the same HGX B300 eight-GPU module in an 8U chassis with onboard ConnectX-8 networking. Cooling and power delivery are properties of the host; plan capacity from the OEM server specification, not from an unpublished per-GPU wattage.

Choose B300 when you need the Blackwell Ultra memory and attention improvements in an eight-GPU NVLink domain. The related B200 180GB SXM remains the documented per-GPU SKU for the preceding HGX B200 board, where NVIDIA does publish 180 GB and a 1,000 W ceiling. Request a quote for B300-class modules or a complete HGX B300 server.